The rendered sample from step 4 is the most representative one. It pulls 4 rows from each of the four `sleep_debt_category` groups, 16 rows in all, so rare categories are not drowned out by the common one. The earlier plain samples (steps 0 and 5) were 15 rows each. **What the sample shows** - **Grain and mix:** each row is one user (`user_id` such as USR-04718), with demographics (age, gender, occupation, chronotype), bedtime behaviour (app, screen brightness, blue-light filter, caffeine after 5pm, physical activity), and sleep outcomes. - **Category skew:** across all 8,500 rows, Moderate Debt is 52.5% (4,462 users), Mild Deficit 23.6% (2,004), Optimal Recovery 16.3% (1,387) and Severe Sleep Debt 7.6% (647). A purely random sample would mostly show Moderate Debt, which is why I stratified. - **Gradient across categories:** the sample rows reflect the group averages. Optimal Recovery averages 8.35 sleep hours, 28.9 minutes of sleep latency, 0.81 alarm snoozes and a fatigue score of 1.08. Severe Sleep Debt averages 3.76 hours, 72.7 minutes of latency, 5.88 snoozes and a fatigue score of 9.58. For example, USR-06860 in the sample has 3.2 hours of sleep, 75 minutes of latency, 7 snoozes and a fatigue score of 10.0. - **Typical value patterns:** - Many users report 0 mg of caffeine after 5pm, and the median is 0 in every category except Severe Sleep Debt, where it is 25 mg. - `blue_light_filter_active` is a 0/1 flag. - Ages in the sample range from 20 to 65, matching the dataset range of 18–65. - Severe Sleep Debt rows tend to have higher screen brightness (average 60.8% versus about 54–55% in the other groups). **Notes** - `bedtime_phone_minutes` is flagged as person-identifying, so I left it out of the row-level samples. Its category averages are available from the aggregate query (step 10). - The audit columns (`_ingestion_timestamp`, `_batch_id`, `_source_file`, `_source_system`) are also left out because they carry no analytical content. - The samples are small and illustrative, not statistical evidence. The category averages come from the full 8,500 rows. I can pull a larger or differently stratified sample, for example by occupation or chronotype, if that would help.